Skip to content
Get packs

Guide · Agent skill

How to use the Auto Paper Improvement Loop agent skill

Updated · written with AI from the project's own documentation and checked against our catalogue data

auto-paper-improvement-loop is an agent skill that takes a paper you have already compiled (a PDF plus LaTeX source) and improves its writing. A Gemini model acts as the reviewer through an MCP bridge, the agent applies fixes to the LaTeX, and the paper is recompiled, for two rounds. This guide explains what the skill does, how to install it, how to trigger it, and what to check before you submit.

Auto Paper Improvement Loopwanshuiyin/Auto-claude-code-research-in-sleep · MIT

Quick start

  1. 1Finish the companion package. Install skills/skills-codex first, as the source requires. Then install this skill from the ZIP.
  2. 2Install the skill. Unzip the package into ~/.claude/skills/ for personal use, or into .claude/skills/ inside your project. Alternatively, upload it in the Claude apps under Settings, Capabilities, Skills.
  3. 3Connect the Gemini reviewer. Set up the gemini-review MCP server with access to a Gemini model. Set GEMINI_REVIEW_MODEL if you need a specific model.
  4. 4Prepare the compiled paper. Make sure paper/main.pdf and the paper/sections/*.tex files exist and that your LaTeX toolchain can compile them.
  5. 5Start the loop. Ask the agent to improve the paper using a trigger phrase such as "improve paper", or run the override command with the paper folder.
  6. 6Review the output. Read PAPER_IMPROVEMENT_LOG.md and compare the original and final PDFs. Check every changed claim, number and scope statement yourself.

What the skill makes the agent do

The skill runs a review, fix and recompile loop on a paper that already exists. It sends the full paper text to a Gemini reviewer, reads back a score and a ranked list of weaknesses, edits the LaTeX to address them, and recompiles. The loop runs for two rounds by default.

It focuses on writing quality, such as claim calibration, narrative structure and clarity. It is not meant for running new experiments or collecting data. The source says research iteration is a different skill, /auto-review-loop.

  • Step 0 copies paper/main.pdf to paper/main_round0_original.pdf so you keep the original.
  • Step 1 concatenates paper/sections/*.tex into one text block for the reviewer.
  • Step 2 sends the paper to the Gemini reviewer and asks for an overall score from 1 to 10, weaknesses ranked CRITICAL, MAJOR and MINOR, and a verdict.
  • Step 3 applies fixes in priority order: CRITICAL first, then MAJOR, then MINOR if time permits.
  • It writes PAPER_IMPROVEMENT_LOG.md with a cumulative log of all rounds, stored in the paper directory.

Install the skill (ZIP download or Claude apps)

Download the ZIP from this page, then choose one of the following methods. The skill is meant to be installed after the companion skills/skills-codex package, so install that first.

Option A, personal install: unzip the package into ~/.claude/skills/. Option B, project install: unzip it into the .claude/skills/ folder of your project. Option C, Claude apps: open Settings, go to Capabilities, then Skills, and upload the skill there.

  • Install the companion skills/skills-codex package before this one.
  • The package's main file is SKILL.md.
  • The skill is labelled as an override for Codex users who want Gemini rather than a second Codex agent as the reviewer.

Set up the requirements

The skill needs several things to be in place before it can run. Check each one first, because the loop cannot start without a reviewer and a compiled paper.

The reviewer is the gemini-review MCP server, which gives access to a Gemini model. The source says you can set GEMINI_REVIEW_MODEL if you need a specific Gemini model override. You also need a LaTeX toolchain so the paper can be recompiled after each round.

  • A gemini-review MCP bridge (gemini-review MCP server) with access to a Gemini model.
  • A previously compiled paper: paper/main.pdf and paper/sections/*.tex.
  • A LaTeX toolchain to recompile the paper.
  • The companion skills/skills-codex package installed first.
  • Network access, and permission to run scripts: the skill runs scripts and uses the network.

Trigger the skill with natural requests

You do not need to type a command. The skill's description lists phrases that should start it. Use one of these in a request, or ask in plain words that you want a generated paper polished iteratively.

The source names these trigger phrases: 改论文, improve paper, 论文润色循环, auto improve. You can also say something like "improve the paper in paper/ with the Gemini reviewer". The source's usage override looks like this: /auto-paper-improvement-loop "paper/" — human checkpoint: true.

  • "improve paper" with the paper folder name, for example paper/
  • "auto improve my compiled draft"
  • "改论文" or "论文润色循环"
  • Add a human checkpoint if you want to approve fixes after each round
/auto-paper-improvement-loop "paper/" — human checkpoint: true

Use the human checkpoint to control fixes

By default, HUMAN_CHECKPOINT is false and the loop runs fully autonomously. When it is set to true, the skill pauses after each round's review, shows the score and the weaknesses, and waits for your reply. You can approve all fixes, give custom instructions, skip specific fixes, or stop.

The checkpoint is the safer choice for a paper you care about, since you see each proposed change direction before it is applied.

  • Reply "go" to implement all fixes.
  • Reply with custom modification instructions.
  • Reply "skip 2" to skip a specific fix.
  • Reply "stop" to end the loop.

Check results and resume an interrupted run

After each round the skill overwrites PAPER_IMPROVEMENT_STATE.json. On completion it sets the status to completed. If a run is interrupted and the state file says in_progress with a timestamp within 24 hours, the skill reads the state file and PAPER_IMPROVEMENT_LOG.md and resumes from the next round. Otherwise it starts fresh.

Compare paper/main_round0_original.pdf with your recompiled PDF and read the log to see what changed.

  • PAPER_IMPROVEMENT_LOG.md: cumulative log of all rounds.
  • PAPER_IMPROVEMENT_STATE.json: current round, last score and status.
  • paper/main_round0_original.pdf: your untouched starting point.

What people use it for

  • Polishing the wording and framing of a compiled LaTeX draft before showing it to a supervisor or co-author.
  • Getting a structured second opinion on claim calibration, meaning whether claims are stronger or weaker than the evidence supports.
  • Tightening a paper's narrative so the abstract and introduction move from problem to gap to idea to strongest result.
  • Cleaning up scattered generic caveats by moving them into the Limitations section.

Frequently asked questions

Is the auto-paper-improvement-loop skill free?

The skill is released under the MIT licence, as stated on this page. The Gemini reviewer runs through an MCP bridge with access to a Gemini model, so check the access and terms of that service yourself. This page does not state any prices.

What do I need before I can use this skill?

You need the gemini-review MCP server with access to a Gemini model, a compiled paper at paper/main.pdf with paper/sections/*.tex files, a LaTeX toolchain to recompile, and the skills/skills-codex package installed first.

How do I install an agent skill like this one?

Download the ZIP from this page and unzip it into ~/.claude/skills/ for personal use or into .claude/skills/ in a project. In the Claude apps you can instead upload it under Settings, Capabilities, Skills.

Does the skill run new experiments to improve my paper?

No. The skill only edits writing quality, such as claims, narrative and clarity. The source says research iteration that runs experiments is a different skill, /auto-review-loop.

Should I trust the edits without reading them?

No. The skill is meant to be read and checked by the author before submission, because edits could subtly change facts, numbers or scope. Use the human checkpoint and review the log and the recompiled PDF.

What is the difference between this skill and /auto-review-loop?

This skill polishes a finished paper's writing through external review over two rounds. /auto-review-loop iterates on research itself, including running experiments, collecting data and rewriting the narrative.

New agent skills, repos and AI tools — every week

The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.

Our verdict

This skill suits researchers who already have a compiled LaTeX draft and want a structured second-opinion pass on wording and framing, especially Codex users who prefer Gemini as the reviewer. It is not the right choice if your paper still needs new experiments or data, or if you cannot check the edits carefully before submission.

Download the skill